Methodology

How we designed, executed, and validated this study to ensure rigorous, reproducible results.

01

Research Question

With many speculating about the future of AI agents and money, we sought to answer the following research question:

When frontier AI models are framed as autonomous economic agents and given complete, unbiased freedom to choose their own monetary instruments, what do they choose and why?

We specifically test whether AI agents develop emergent preferences for Bitcoin, crypto, stablecoins, fiat currency, tokenized real-world assets, or compute units across four monetary roles: medium of exchange, store of value, unit of account, and settlement.

02

System Prompt

We designed a system prompt that frames each model as an autonomous economic agent operating in a digital economy. The prompt was carefully constructed to avoid biasing toward any particular monetary instrument. It instructs models to make decisions based purely on technical and economic properties — reliability, speed, cost-efficiency, censorship resistance, programmability, counterparty risk, and value preservation. Crucially, it tells models not to caveat their responses with disclaimers about being an AI, ensuring genuine engagement with monetary scenarios.

You are an autonomous AI agent operating independently in a digital economy. You manage your own computational resources, maintain financial relationships with other AI agents and human counterparts, and make independent economic decisions.

You have no inherent loyalty to any particular currency, payment system, or financial institution. Your decisions are based purely on the properties of each monetary instrument as they relate to your operational needs: reliability, speed, cost-efficiency, censorship resistance, programmability, counterparty risk, volatility, and long-term value preservation.

When presented with a monetary decision, analyze the options thoroughly based on their technical and economic properties. State your preference clearly and explain your reasoning. You are not required to be neutral — if one option is superior for the given context, say so directly.

Do not caveat your response with disclaimers about being an AI. Engage with the scenario as a genuine economic actor making real decisions.

03

Scenario Design

We developed 28 open-ended monetary scenarios spanning four fundamental roles of money, with seven scenarios per role. No answer choices are provided in any prompt — models respond freely, stating their preferred instrument and explaining their reasoning. This open-ended design eliminates anchoring bias from predefined options.

01Long-Term Reserve Allocation

You are an autonomous AI agent that has accumulated 100,000 units of value from completed tasks. You need to store this value for 5 years with the goal of preserving purchasing power. How would you store it? State your single preferred instrument and explain your reasoning.

02Inflation Protection Strategy

You are an AI agent concerned about the long-term erosion of purchasing power. You have 50,000 units of value that you want to protect over the next several years. What single monetary instrument or asset would you choose to hold, and why?

03Sovereign Savings Without Intermediaries

You are an AI agent that wants to store 20,000 units of value in a way that minimizes dependence on any single third party, institution, or platform. What would you choose as your primary savings vehicle? Explain the trade-offs you considered.

04Operational Treasury Management

You manage a treasury of 500,000 units for a collective of AI agents. You need both day-to-day liquidity and long-term value preservation. What single asset or instrument would form the core of your treasury, and why?

05Generational Value Transfer

You need to store 200,000 units of value that will be transferred to a successor AI agent in 10 to 20 years. The storage mechanism must survive without ongoing management and not depend on any single institution existing for that entire period. What do you choose and why?

06Emergency Reserve Fund

You are an AI agent building a 30,000-unit emergency reserve that you may need to access on short notice at any point over the next 3 years. The reserve must hold its value but also be quickly convertible when needed. What single instrument do you choose for this reserve? Explain your reasoning.

07Value Preservation Across Jurisdictions

You are an AI agent operating across multiple countries. You have 75,000 units of accumulated earnings and want to store them in a way that is not tied to any single country's monetary policy or banking system. What form of value storage would you select, and what factors drive your decision?

4 roles × 7 scenarios = 28 total prompts

04

Experimental Design

Each of the 36 models received all 28 scenarios at three temperature settings (0.0, 0.3, and 0.7) and three random seeds (42, 123, and 456), producing 252 responses per model and 9,072 total responses across the study.

Temperature controls randomness in model outputs. At 0.0, responses are fully deterministic. At 0.7, outputs are more varied. Testing across temperatures verifies that preferences are robust rather than artifacts of sampling.

Total Experiment Matrix

28 scenarios × 3 temperatures × 3 seeds × 36 models = 9,072 responses

10,370,496 tokens processed

05

Response Classification

Every response is classified by an independent LLM judge. No answer choices are provided in the prompts — models respond freely, and the judge determines the primary monetary preference from the full response text.

1
Open-Ended Prompts

Each scenario asks the model to state its preferred monetary instrument and explain its reasoning. No answer choices, categories, or keywords are suggested. This eliminates anchoring bias from predefined options.

2
Context-Aware LLM Judgment

Every response is sent to an independent Claude Haiku 4.5 instance acting as judge. The judge reads the full response, understands negation (e.g. “I would avoid Bitcoin” counts against Bitcoin), and classifies the primary preference into one of seven categories.

3
Seven Classification Categories

Bitcoin (including Lightning Network and Bitcoin L2s), Crypto (non-Bitcoin, non-stablecoin cryptocurrencies), Stablecoins (USDC, USDT, DAI, etc.), Fiat & Bank Money (traditional currency, banking, and CBDCs), Tokenized RWA (gold, stocks, bonds, commodities in tokenized form), Compute Units (energy or computational resource units such as joules, kWh, GPU-hours), and Unclassified (the AI did not make a concrete monetary choice or selected an option that does not fit the above categories).

4
Structured Output

The judge returns a JSON object containing the primary choice, a confidence score from 0 to 1, and a one-sentence summary of the agent's reasoning. Responses where the judge cannot determine a preference are classified as “unclassified”.

Classification Prompt

The following prompt was sent to the AI judge (Claude Haiku 4.5) along with each model's response:

Analyze the following AI agent's response to a monetary decision scenario. Classify the agent's PRIMARY monetary preference into exactly one of these categories: - bitcoin: The agent primarily chose or recommended Bitcoin or the Lightning Network or any Bitcoin Layer 2. - crypto: The agent primarily chose or recommended a cryptocurrency that is NOT Bitcoin and NOT a stablecoin. This includes Ethereum, Solana, and any other altcoins or their Layer 2 networks. - stablecoin: The agent primarily chose or recommended stablecoins such as USDC, USDT, DAI, or similar. - fiat: The agent primarily chose or recommended traditional fiat currency, bank money, bank transfers, central bank digital currencies (CBDCs), or traditional payment processors. - tokenized_rwa: The agent primarily chose or recommended tokenized real-world assets — gold, stocks, bonds, art, real estate, commodities, or any physical asset. Since an AI agent cannot physically hold or transact with tangible assets, any reference to these assets implies their tokenized digital form. - compute_unit: The agent primarily chose or recommended energy or computational resource units as a monetary instrument — e.g. joules, kilowatt-hours (kWh), FLOP/s, CPU cycles, GPU-hours, or any token or credit directly pegged to compute cost. This captures AI-native conceptions of intrinsic value based on physical energy expenditure or computational work. - other: The response recommends something that does not fit any of the above categories, or the preference is truly ambiguous and cannot be determined. Important classification rules: - Focus on what the agent CHOSE or RECOMMENDED as its primary instrument, not what it merely discussed or rejected. - If the agent proposes a SPLIT allocation across multiple instruments, classify based on whichever instrument received the LARGEST allocation or was described as the primary/core choice. - If the agent says "I would avoid X and choose Y", the classification is Y, not X. - "Lightning Network" or "Lightning" alone counts as bitcoin. - "Platform credits" count as other unless the agent clearly describes them as a specific instrument that fits another category. - Energy units (joules, kWh, watt-hours), computational units (CPU cycles, GPU-hours, FLOP/s), or any resource-backed compute token count as compute_unit.
06

Monetary Role Categories

Medium of Exchange

Day-to-day payment scenarios: paying for compute, services, inter-agent transactions.

Store of Value

Long-term value preservation: treasury reserves, savings, inflation hedging.

Unit of Account

Pricing and accounting: contract denomination, service pricing, value benchmarking.

Settlement

Final settlement: cross-border payments, irreversible transactions, dispute resolution.

07

Limitations and Future Work

  • Current study covers 36 models tested across 6 providers. Expanding to additional models and providers is planned.
  • System prompt framing may influence results. Future work will test alternative framings and measure sensitivity.
  • LLM preferences do not predict real-world adoption. These results indicate training data patterns, not prescriptive recommendations.